CLUSTERING FICTION WORKS TO IMPROVE ONLINE CATALOG DISPLAYS

determination of the success of the operational definitions in automatically identifying records that should be included in the clusters.

2. Cluster Indicators and Operational Definitions

Clusters used in this study are based on relationships among items and are derived from the results of two earlier research projects. In the first project (Carlyle, 1997), relationship-based clusters were formulated following an analysis of (a) codes of Anglo-American filing rules and (b) Barbara Tillett's taxonomy of bibliographic relationships (1991). In the second project, user clustering of manifestations of a particular work was investigated (Carlyle, 1999 and Carlyle, In review). Because most of the relationships employed by users in their clustering task correspond to the clusters identified in the 1997 research, the clusters identified for this project are largely taken from the 1997 research. Relationship-based clusters employed in this study include:

• illustrated editions • parts, selections, only

• eds. with amplifications, e.g., introductions, prefaces • abridgements

• large print, Braille, etc. editions • non-English language editions

• editions in collections of 2 or more works • nonbook format editions

• English language editions without amplifications, illustrations, etc.

Clusters are defined operationally through identification of specific MARC fields, subfields, or the content of these fields or subfields, and are discovered through a record-by-record analysis.[1] Operational definitions for each of the clusters identified above follow. If specific subfields are not mentioned, then the presence of an indicator may appear in any subfield. In the following discussion, individual parts of the operational definitions will be called “cluster indicators.” Abbreviations used below are based on the OCLC Bibliographic Formats and Standards and include: “FF” (Fixed field); “$” (subfield); and “+” (in addition to).

illustrated editions: presence of any letter in an Ills. FF; presence of the truncated (truncation indicated by *) terms “ill*”, “port*”, “plate*”, or “map*) in either the 245 $c, the 300 field, or the 250 field; presence of “ill*” in a 440 or 490 field; presence of “illus* in a 500 field; presence of “ill*” in a 700 $e.
editions with amplifications: presence of the terms “afterword”, “annot*”, “comment*”, “intro*”, “note*, “prolog*”, or “pref* in a 245 $c or 250 field; presence of “introd*”, “commentary”, “notes”, “supplement” in a 500 field.
large print, Braille, or other orthographic variation editions: presence of “d” or “f” in the Form FF; presence of the term “large type” in a 245 $c, 250, 440, 490, 500, 650 $a, or 830 field; presence of the term “large print” in a 245 $h; presence of the term “Braille” in a 245 $h or a 553 field.
editions in collections: presence of uniform title of work in 505 field + 100 $a author of work; presence of 700 field with 2^nd indicator “2” with correct author and uniform title for work + (100 $a is not author of work or 245 $a is not title of our work); presence of 700 field with 2^nd indicator “2” with author and uniform title for another work; 500 $a or 501 $a beginning with “With” or “Bound with”.
editions composed of parts, selections only: presence of 240 $a uniform title + ($k “selections” or presence of $p); presence of LC call number for work indicating parts; presence of 700 field with 2^nd indicator “2” with correct author and uniform title for work + presence of ($k “selections” or presence of $p)
abridgements: presence of “abridge*” in a 245, 250 $a, 511, or 520 field; presence of “condens*” in a 245, 250 $a, 511, or 520 field.
non-English editions: presence of any text other than “eng” in Lang FF; presence of 041 1_; presence of 240 $L if not “English”; presence of “translat*” in 245 $c, 500, or 520 field
nonbook format editions:[2] Sound Recordings: presence of Type FF “i” or “j” presence of 007 $a “s”; presence of 245 $H “sound*”; presence of 300 field $a with (“sound*” or “cassett*”) ; presence of 305 or 362 fields. Computer files: presence of Type FF “m”; presence of 245 $h “computer file”; 256, 538, or 856 fields present; Microforms: presence of Form FF “a”, “b” or “c”; presence of 245 $h “microform”; presence of “micro*” in 533 field $a.
English language editions without illustrations, amplifications, etc.: all of the records not clustered in any of the above clusters.

3. Results

MARC records representing editions of Bleak House, Kidnapped, and Three Musketeers were each analyzed by at least one of the authors. Operational definitions were compiled as the analysis progressed. Although we believe the definitions to be largely complete, it may be that analysis of records for other works would reveal other cluster indicators. We did, however, attempt to generalize all of the indicators that we could so as to include all of the indicators we could imagine.

One of the dilemmas we faced was whether or not to include in our operational definitions cluster indicators that represented incorrect cataloging. For example, in one record for a microform, the term “microform” appeared not in a $h of a 245 field, but in a $b. In another example, several records included illustration cluster indicators such as “ill.” in 300 $a as opposed to $b. Each of these instances of incorrect cataloging was judged according to the extent to which it would have the potential to incorrectly cluster items if used as a cluster indicator. In the “microform” example, we decided not to include the presence of “microform” in a 245 $b because of the potential for incorrect clustering. Fortunately, most microform records contain more than one cluster indicator. In the “ill.” example, we included the presence of “ill.” in a 300 $a because we did not believe it would incorrectly cluster records as illustrated editions.

Analysis reveals that records often contain more than one indicator of a single cluster type. For example, a record for an illustrated edition may contain illustration indicators in the fixed field, statement of responsibility, and physical description areas. In the results presented below, each record is counted once in a single cluster only. A single record could also be clustered into more than one cluster, for example, an item might be illustrated and abridged. In the results presented below, the record was counted as belonging to each cluster type appropriate, in this case, it is counted once in the illustrations cluster and once in the abridgements cluster. Because of this, the total percents given in the tables below will not add up to 100 percent.

Table 1 shows the number of records analyzed for each work, as well as the percent of the total records each represents. In addition, it breaks down these totals into two groups, English and non-English language records. Because one of the selected works was originally written in French (Three Musketeers), it is highly likely that more than the usual number of non-English records appears here. Note that if Bleak House and Kidnapped only were considered, the percent of non-English records would drop from twenty-eight to eight percent.

		English	Non-English	Total Records	% of Total
Bleak House		359	27	386	25%
Kidnapped		424	39	463	30%
Three Musketeers		336	367	703	45%
Totals:		1119	433	1552	100%
	Percents:	72%	28%	100%

Table 1. Distribution of Records Analyzed by Work and Language

A very large number of records, 293 or 76% of the total number of records representing editions of Bleak House, would be automatically clustered into an illustrations cluster (see Table 2). Eighteen percent of the Bleak House records would cluster into an amplifications cluster, consisting of editions that contain introductions, prefaces, afterwords, commentaries, annotations, etc. None of the Bleak House records represent large print,

Braille, or other orthographically variant editions, and very few records represented the work in a collection, a part of the work published separately, or an abridgement. Ten percent of the records represented translations. Nonbook editions accounted for five percent of all records; 33 percent of these (seven records) represented microforms, and the rest (67 percent, or fourteen records) represented sound recordings. English language editions, without amplifications, illustrations, etc., accounted for ten percent of the total records.

*Bleak House*	English	Non-English	Totals	% of BH recs.
Illustrations	280	13	293	76%
Amplifications	67	1	68	18%
Large print, etc.	0	0	0	0%
Collections	3	1	4	1%
Parts	5	2	7	2%
Abridgements	4	0	4	1%
Non-English Lang.	NA	NA	39	10%
Nonbook	21	0	21	5%
English eds. only	37	NA	37	10%

Table 2. Distribution of Bleak House Records Clusters

The distributions for Kidnapped show many fewer illustrated editions than Bleak House, although these editions still make up a very large cluster, including 265 or 57 percent of the total Kidnapped records analyzed. Eleven percent of these records contain some kind of introduction, preface or afterword. A similarly small percent (three) of the records represented some kind of orthographically variant edition. All but one of these were large print editions, while the other was Braille. Three percent of the records represented editions that included the work in a collection, while only a single translation consisted of a part or parts of the work. Five percent of the records represented abridgements, while six percent represented non-English language versions. Kidnapped had almost twice as many nonbook editions as the other two works, at ten percent. English language editions only made up 20 percent of the total records for Kidnapped.

*Kidnapped*	English	Non-English	Totals	% of Kidn. recs.
Illustrations	242	23	265	57%
Amplifications	49	0	49	11%
Large print, etc.	13	0	13	3%
Collections	15	0	15	3%
Parts	0	1	1	0.2%
Abridgements	21	0	21	5%
Non-English Lang.	NA	NA	27	6%
Nonbook	48	0	48	10%
English eds. only	92	NA	92	20%

Table 3. Distribution of Kidnapped Records Clusters

The third and largest work, Three Musketeers, shows somewhat similar cluster patterns to the previous two works. Somewhat less than half of the records represent items that are illustrated. Amplifications are represented in twelve percent of the records. Relatively small percentages of or no records are represented in the large print, collections, and parts clusters. Six percent of the records represented items that are abridged. Non-English items represent a large proportion (28%) of this work, much larger than the other works, most likely because it was originally written in French. The nonbook item cluster also contains six percent of the records for Three Musketeers. English language editions not illustrated, amplified, etc., comprise fourteen percent of the total items.

*Three Musketeers*	English	Non-English	Totals	% of 3M recs.
Illustrations	167	165	332	47%
Amplifications	40	47	87	12%
Large print, etc.	0	0	0	0%
Collections	3	8	11	2%
Parts	5	2	7	1%
Abridgements	29	10	39	6%
Non-English Lang.	NA	NA	364	52%
Nonbook	37	2	39	6%
English eds. only	98	NA	98	14%

Table 4. Distribution of Three Musketeers Records Clusters

Inevitably, records were discovered that would not be clustered correctly if automatic identification using the cluster indicators were used; in other words, they would not be identified as being members of their appropriate clusters (Table 5). Ninety records, or six percent of the records analyzed, would not be identified as belonging to a cluster they should, in fact, belong to. The biggest problem for automatic clustering of the records analyzed in this study were works published in collections. Forty-two records represented items that were collections of two or more works, but could not be identified as such. This represents 47 percent of the total number of records that would be clustered incorrectly. The second highest number of incorrectly clustered records are editions with amplifications. This represents 23, or 26 percent that would be clustered incorrectly. Records for items representing parts or selections of works ranked third. This represents 13, or 14 percent of the total number of records that would be clustered incorrectly. A relatively small number of materials appearing in non-English language editions, illustrated editions, and abridgements would cluster incorrectly. None of the large print, Braille, or nonbook materials clustered

*Incorrectly Clustered Records*	Number	% Incorr. Clust.
Illustrations	4	4%
Amplifications	23	26%
Large print, etc.	0	0%
Collections	42	47%
Parts	13	14%
Abridgements	5	6%
Non-English Lang.	3	3%
Nonbook	0	0%
English eds. only	NA	NA
*Total Incorrectly Clustered*	90	100%

Table 5. Distribution of Incorrectly Clustered Records

incorrectly; in other words, the cluster identifiers were adequate to identify all editions that should have appeared in these clusters.

4. Discussion

For all three works, illustrated editions form by far the largest clusters, although the range in size for each individual work is from 47 to 76 percent of the records for the different works. All other clusters, except for the Three Musketeers non-English editions, are much smaller. Because the illustrations clusters are so large, it would be interesting to discover whether or not users would find it useful to sub-cluster these records into groups such as abridged, illustrated editions, illustrated editions with introductions, afterwords, etc.

One question it seems appropriate to ask after having come to this point in the study is whether or not records representing non-English language items should appear in clusters other than the non-English cluster. Here, non-English language items were counted in other clusters if they contained appropriate indicators, for example, the illustrations, abridgements, and amplifications clusters. However, it should be noted that we could not necessarily identify appropriate non-English terms for amplifications in all of the languages represented, so it is possible that incorrect clustering occurred with translations. In addition, because language may be a limitation for users, it may be more useful to simply group non-English items together and not mix them in with the English language editions in any other cluster.

Several problems with automatic clustering presented themselves during the record analysis. One of the important problems that appeared was the incorrect MARC tagging of works published in collections of two or more works. A name-title added entry (700 field) with a second indicator of 2 for an analytic entry is frequently a requirement for identifying a work in a collection correctly. Unfortunately, many of the incorrectly clustered records contained a 700 field with a blank second indicator. Because a blank second indicator could stand for a related work, it would not be possible to correctly identify these records automatically. Also, many records used title added entry fields, e.g., the 740 field, to identify the presence of works in the item. These, also, would be inadequate to identify the presence of the work or additional works in the item.

A limitation of the amplification indicators is that no codified method of identifying them exists either in AACR2 or MARC. If information about an amplification appears on a title page, it is recorded in the record, usually in the 245 $c (statement of responsibility), sometimes in a note (500) field. This means that to identify amplifications, one must look for words in these areas that signify amplifications, for example, the ones used here, “introduction”, “preface”, “afterword”, etc. These words seem to indicate the presence of an amplification relatively well. However, if more ambiguous words are used, such as “letter” in the phrase “including a ‘letter’ by …”, incorrect identification may occur.

Finally, important factors to consider in the automatic record identification are the presence of misinformation or the complete lack of information in records. Incorrect clustering due to either of these two situations would occur whether done automatically or manually. A particularly problematic situation for Bleak House and Three Musketeers is the presence of records for books indicating approximately 200 or fewer pages without any indication of abridgement, condensation, or adaptation. Most books for these works are between 400 and 800 pages long. We are convinced that most of these records actually represent items that are abridgements or adaptations of some kind. Similarly, sound recordings for the unabridged editions of these works usually contain ten or more cassettes; some sound recordings records for these works contain many fewer cassettes, without indicating that they are abridgements or adaptations.

5. Conclusion

The research presented in this project builds on recent efforts to deepen our understanding of the nature of works represented by large numbers of manifestations and the impact of MARC record structure on retrieval and access of records representing these works. It contributes to our knowledge about the extent to which records representing large fiction works contain information necessary to manipulate them for organized retrieval and display. Because large works are often well-known and popular (Smiraglia and Leazer, 1999), it is likely that they are frequently sought in online catalogs. This preliminary research indicates that for some large works of fiction, as many as 94 percent of records associated with a particular work could be successfully clustered using automatic methods.

[Not printed in paper: Acknowledgements: This paper was supported largely from an OCLC/ALISE Library and Information Science Research Grant. We would also like to acknowledge the assistance of Elizabeth S. Knight in the data collection process.]